drivers/md/persistent-data/dm-btree.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011 Red Hat, Inc.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include "dm-btree-internal.h"
   9 #include "dm-space-map.h"
  10 #include "dm-transaction-manager.h"
  11
  12 #include <linux/export.h>
  13 #include <linux/device-mapper.h>
  14
  15 #define DM_MSG_PREFIX "btree"
  16
  17 /*
  18  *--------------------------------------------------------------
  19  * Array manipulation
  20  *--------------------------------------------------------------
  21  */
  22 static void memcpy_disk(void *dest, const void *src, size_t len)
  23         __dm_written_to_disk(src)
  24 {
  25         memcpy(dest, src, len);
  26         __dm_unbless_for_disk(src);
  27 }
  28
  29 static void array_insert(void *base, size_t elt_size, unsigned int nr_elts,
  30                          unsigned int index, void *elt)
  31         __dm_written_to_disk(elt)
  32 {
  33         if (index < nr_elts)
  34                 memmove(base + (elt_size * (index + 1)),
  35                         base + (elt_size * index),
  36                         (nr_elts - index) * elt_size);
  37
  38         memcpy_disk(base + (elt_size * index), elt, elt_size);
  39 }
  40
  41 /*----------------------------------------------------------------*/
  42
  43 /* makes the assumption that no two keys are the same. */
  44 static int bsearch(struct btree_node *n, uint64_t key, int want_hi)
  45 {
  46         int lo = -1, hi = le32_to_cpu(n->header.nr_entries);
  47
  48         while (hi - lo > 1) {
  49                 int mid = lo + ((hi - lo) / 2);
  50                 uint64_t mid_key = le64_to_cpu(n->keys[mid]);
  51
  52                 if (mid_key == key)
  53                         return mid;
  54
  55                 if (mid_key < key)
  56                         lo = mid;
  57                 else
  58                         hi = mid;
  59         }
  60
  61         return want_hi ? hi : lo;
  62 }
  63
  64 int lower_bound(struct btree_node *n, uint64_t key)
  65 {
  66         return bsearch(n, key, 0);
  67 }
  68
  69 static int upper_bound(struct btree_node *n, uint64_t key)
  70 {
  71         return bsearch(n, key, 1);
  72 }
  73
  74 void inc_children(struct dm_transaction_manager *tm, struct btree_node *n,
  75                   struct dm_btree_value_type *vt)
  76 {
  77         uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
  78
  79         if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
  80                 dm_tm_with_runs(tm, value_ptr(n, 0), nr_entries, dm_tm_inc_range);
  81
  82         else if (vt->inc)
  83                 vt->inc(vt->context, value_ptr(n, 0), nr_entries);
  84 }
  85
  86 static int insert_at(size_t value_size, struct btree_node *node, unsigned int index,
  87                      uint64_t key, void *value)
  88         __dm_written_to_disk(value)
  89 {
  90         uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
  91         uint32_t max_entries = le32_to_cpu(node->header.max_entries);
  92         __le64 key_le = cpu_to_le64(key);
  93
  94         if (index > nr_entries ||
  95             index >= max_entries ||
  96             nr_entries >= max_entries) {
  97                 DMERR("too many entries in btree node for insert");
  98                 __dm_unbless_for_disk(value);
  99                 return -ENOMEM;
 100         }
 101
 102         __dm_bless_for_disk(&key_le);
 103
 104         array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
 105         array_insert(value_base(node), value_size, nr_entries, index, value);
 106         node->header.nr_entries = cpu_to_le32(nr_entries + 1);
 107
 108         return 0;
 109 }
 110
 111 /*----------------------------------------------------------------*/
 112
 113 /*
 114  * We want 3n entries (for some n).  This works more nicely for repeated
 115  * insert remove loops than (2n + 1).
 116  */
 117 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
 118 {
 119         uint32_t total, n;
 120         size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
 121
 122         block_size -= sizeof(struct node_header);
 123         total = block_size / elt_size;
 124         n = total / 3;          /* rounds down */
 125
 126         return 3 * n;
 127 }
 128
 129 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
 130 {
 131         int r;
 132         struct dm_block *b;
 133         struct btree_node *n;
 134         size_t block_size;
 135         uint32_t max_entries;
 136
 137         r = new_block(info, &b);
 138         if (r < 0)
 139                 return r;
 140
 141         block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
 142         max_entries = calc_max_entries(info->value_type.size, block_size);
 143
 144         n = dm_block_data(b);
 145         memset(n, 0, block_size);
 146         n->header.flags = cpu_to_le32(LEAF_NODE);
 147         n->header.nr_entries = cpu_to_le32(0);
 148         n->header.max_entries = cpu_to_le32(max_entries);
 149         n->header.value_size = cpu_to_le32(info->value_type.size);
 150
 151         *root = dm_block_location(b);
 152         unlock_block(info, b);
 153
 154         return 0;
 155 }
 156 EXPORT_SYMBOL_GPL(dm_btree_empty);
 157
 158 /*----------------------------------------------------------------*/
 159
 160 /*
 161  * Deletion uses a recursive algorithm, since we have limited stack space
 162  * we explicitly manage our own stack on the heap.
 163  */
 164 #define MAX_SPINE_DEPTH 64
 165 struct frame {
 166         struct dm_block *b;
 167         struct btree_node *n;
 168         unsigned int level;
 169         unsigned int nr_children;
 170         unsigned int current_child;
 171 };
 172
 173 struct del_stack {
 174         struct dm_btree_info *info;
 175         struct dm_transaction_manager *tm;
 176         int top;
 177         struct frame spine[MAX_SPINE_DEPTH];
 178 };
 179
 180 static int top_frame(struct del_stack *s, struct frame **f)
 181 {
 182         if (s->top < 0) {
 183                 DMERR("btree deletion stack empty");
 184                 return -EINVAL;
 185         }
 186
 187         *f = s->spine + s->top;
 188
 189         return 0;
 190 }
 191
 192 static int unprocessed_frames(struct del_stack *s)
 193 {
 194         return s->top >= 0;
 195 }
 196
 197 static void prefetch_children(struct del_stack *s, struct frame *f)
 198 {
 199         unsigned int i;
 200         struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
 201
 202         for (i = 0; i < f->nr_children; i++)
 203                 dm_bm_prefetch(bm, value64(f->n, i));
 204 }
 205
 206 static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
 207 {
 208         return f->level < (info->levels - 1);
 209 }
 210
 211 static int push_frame(struct del_stack *s, dm_block_t b, unsigned int level)
 212 {
 213         int r;
 214         uint32_t ref_count;
 215
 216         if (s->top >= MAX_SPINE_DEPTH - 1) {
 217                 DMERR("btree deletion stack out of memory");
 218                 return -ENOMEM;
 219         }
 220
 221         r = dm_tm_ref(s->tm, b, &ref_count);
 222         if (r)
 223                 return r;
 224
 225         if (ref_count > 1)
 226                 /*
 227                  * This is a shared node, so we can just decrement it's
 228                  * reference counter and leave the children.
 229                  */
 230                 dm_tm_dec(s->tm, b);
 231
 232         else {
 233                 uint32_t flags;
 234                 struct frame *f = s->spine + ++s->top;
 235
 236                 r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
 237                 if (r) {
 238                         s->top--;
 239                         return r;
 240                 }
 241
 242                 f->n = dm_block_data(f->b);
 243                 f->level = level;
 244                 f->nr_children = le32_to_cpu(f->n->header.nr_entries);
 245                 f->current_child = 0;
 246
 247                 flags = le32_to_cpu(f->n->header.flags);
 248                 if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
 249                         prefetch_children(s, f);
 250         }
 251
 252         return 0;
 253 }
 254
 255 static void pop_frame(struct del_stack *s)
 256 {
 257         struct frame *f = s->spine + s->top--;
 258
 259         dm_tm_dec(s->tm, dm_block_location(f->b));
 260         dm_tm_unlock(s->tm, f->b);
 261 }
 262
 263 static void unlock_all_frames(struct del_stack *s)
 264 {
 265         struct frame *f;
 266
 267         while (unprocessed_frames(s)) {
 268                 f = s->spine + s->top--;
 269                 dm_tm_unlock(s->tm, f->b);
 270         }
 271 }
 272
 273 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
 274 {
 275         int r;
 276         struct del_stack *s;
 277
 278         /*
 279          * dm_btree_del() is called via an ioctl, as such should be
 280          * considered an FS op.  We can't recurse back into the FS, so we
 281          * allocate GFP_NOFS.
 282          */
 283         s = kmalloc(sizeof(*s), GFP_NOFS);
 284         if (!s)
 285                 return -ENOMEM;
 286         s->info = info;
 287         s->tm = info->tm;
 288         s->top = -1;
 289
 290         r = push_frame(s, root, 0);
 291         if (r)
 292                 goto out;
 293
 294         while (unprocessed_frames(s)) {
 295                 uint32_t flags;
 296                 struct frame *f;
 297                 dm_block_t b;
 298
 299                 r = top_frame(s, &f);
 300                 if (r)
 301                         goto out;
 302
 303                 if (f->current_child >= f->nr_children) {
 304                         pop_frame(s);
 305                         continue;
 306                 }
 307
 308                 flags = le32_to_cpu(f->n->header.flags);
 309                 if (flags & INTERNAL_NODE) {
 310                         b = value64(f->n, f->current_child);
 311                         f->current_child++;
 312                         r = push_frame(s, b, f->level);
 313                         if (r)
 314                                 goto out;
 315
 316                 } else if (is_internal_level(info, f)) {
 317                         b = value64(f->n, f->current_child);
 318                         f->current_child++;
 319                         r = push_frame(s, b, f->level + 1);
 320                         if (r)
 321                                 goto out;
 322
 323                 } else {
 324                         if (info->value_type.dec)
 325                                 info->value_type.dec(info->value_type.context,
 326                                                      value_ptr(f->n, 0), f->nr_children);
 327                         pop_frame(s);
 328                 }
 329         }
 330 out:
 331         if (r) {
 332                 /* cleanup all frames of del_stack */
 333                 unlock_all_frames(s);
 334         }
 335         kfree(s);
 336
 337         return r;
 338 }
 339 EXPORT_SYMBOL_GPL(dm_btree_del);
 340
 341 /*----------------------------------------------------------------*/
 342
 343 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
 344                             int (*search_fn)(struct btree_node *, uint64_t),
 345                             uint64_t *result_key, void *v, size_t value_size)
 346 {
 347         int i, r;
 348         uint32_t flags, nr_entries;
 349
 350         do {
 351                 r = ro_step(s, block);
 352                 if (r < 0)
 353                         return r;
 354
 355                 i = search_fn(ro_node(s), key);
 356
 357                 flags = le32_to_cpu(ro_node(s)->header.flags);
 358                 nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
 359                 if (i < 0 || i >= nr_entries)
 360                         return -ENODATA;
 361
 362                 if (flags & INTERNAL_NODE)
 363                         block = value64(ro_node(s), i);
 364
 365         } while (!(flags & LEAF_NODE));
 366
 367         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 368         if (v)
 369                 memcpy(v, value_ptr(ro_node(s), i), value_size);
 370
 371         return 0;
 372 }
 373
 374 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
 375                     uint64_t *keys, void *value_le)
 376 {
 377         unsigned int level, last_level = info->levels - 1;
 378         int r = -ENODATA;
 379         uint64_t rkey;
 380         __le64 internal_value_le;
 381         struct ro_spine spine;
 382
 383         init_ro_spine(&spine, info);
 384         for (level = 0; level < info->levels; level++) {
 385                 size_t size;
 386                 void *value_p;
 387
 388                 if (level == last_level) {
 389                         value_p = value_le;
 390                         size = info->value_type.size;
 391
 392                 } else {
 393                         value_p = &internal_value_le;
 394                         size = sizeof(uint64_t);
 395                 }
 396
 397                 r = btree_lookup_raw(&spine, root, keys[level],
 398                                      lower_bound, &rkey,
 399                                      value_p, size);
 400
 401                 if (!r) {
 402                         if (rkey != keys[level]) {
 403                                 exit_ro_spine(&spine);
 404                                 return -ENODATA;
 405                         }
 406                 } else {
 407                         exit_ro_spine(&spine);
 408                         return r;
 409                 }
 410
 411                 root = le64_to_cpu(internal_value_le);
 412         }
 413         exit_ro_spine(&spine);
 414
 415         return r;
 416 }
 417 EXPORT_SYMBOL_GPL(dm_btree_lookup);
 418
 419 static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
 420                                        uint64_t key, uint64_t *rkey, void *value_le)
 421 {
 422         int r, i;
 423         uint32_t flags, nr_entries;
 424         struct dm_block *node;
 425         struct btree_node *n;
 426
 427         r = bn_read_lock(info, root, &node);
 428         if (r)
 429                 return r;
 430
 431         n = dm_block_data(node);
 432         flags = le32_to_cpu(n->header.flags);
 433         nr_entries = le32_to_cpu(n->header.nr_entries);
 434
 435         if (flags & INTERNAL_NODE) {
 436                 i = lower_bound(n, key);
 437                 if (i < 0) {
 438                         /*
 439                          * avoid early -ENODATA return when all entries are
 440                          * higher than the search @key.
 441                          */
 442                         i = 0;
 443                 }
 444                 if (i >= nr_entries) {
 445                         r = -ENODATA;
 446                         goto out;
 447                 }
 448
 449                 r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 450                 if (r == -ENODATA && i < (nr_entries - 1)) {
 451                         i++;
 452                         r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 453                 }
 454
 455         } else {
 456                 i = upper_bound(n, key);
 457                 if (i < 0 || i >= nr_entries) {
 458                         r = -ENODATA;
 459                         goto out;
 460                 }
 461
 462                 *rkey = le64_to_cpu(n->keys[i]);
 463                 memcpy(value_le, value_ptr(n, i), info->value_type.size);
 464         }
 465 out:
 466         dm_tm_unlock(info->tm, node);
 467         return r;
 468 }
 469
 470 int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
 471                          uint64_t *keys, uint64_t *rkey, void *value_le)
 472 {
 473         unsigned int level;
 474         int r = -ENODATA;
 475         __le64 internal_value_le;
 476         struct ro_spine spine;
 477
 478         init_ro_spine(&spine, info);
 479         for (level = 0; level < info->levels - 1u; level++) {
 480                 r = btree_lookup_raw(&spine, root, keys[level],
 481                                      lower_bound, rkey,
 482                                      &internal_value_le, sizeof(uint64_t));
 483                 if (r)
 484                         goto out;
 485
 486                 if (*rkey != keys[level]) {
 487                         r = -ENODATA;
 488                         goto out;
 489                 }
 490
 491                 root = le64_to_cpu(internal_value_le);
 492         }
 493
 494         r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
 495 out:
 496         exit_ro_spine(&spine);
 497         return r;
 498 }
 499 EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
 500
 501 /*----------------------------------------------------------------*/
 502
 503 /*
 504  * Copies entries from one region of a btree node to another.  The regions
 505  * must not overlap.
 506  */
 507 static void copy_entries(struct btree_node *dest, unsigned int dest_offset,
 508                          struct btree_node *src, unsigned int src_offset,
 509                          unsigned int count)
 510 {
 511         size_t value_size = le32_to_cpu(dest->header.value_size);
 512
 513         memcpy(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
 514         memcpy(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
 515 }
 516
 517 /*
 518  * Moves entries from one region fo a btree node to another.  The regions
 519  * may overlap.
 520  */
 521 static void move_entries(struct btree_node *dest, unsigned int dest_offset,
 522                          struct btree_node *src, unsigned int src_offset,
 523                          unsigned int count)
 524 {
 525         size_t value_size = le32_to_cpu(dest->header.value_size);
 526
 527         memmove(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
 528         memmove(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
 529 }
 530
 531 /*
 532  * Erases the first 'count' entries of a btree node, shifting following
 533  * entries down into their place.
 534  */
 535 static void shift_down(struct btree_node *n, unsigned int count)
 536 {
 537         move_entries(n, 0, n, count, le32_to_cpu(n->header.nr_entries) - count);
 538 }
 539
 540 /*
 541  * Moves entries in a btree node up 'count' places, making space for
 542  * new entries at the start of the node.
 543  */
 544 static void shift_up(struct btree_node *n, unsigned int count)
 545 {
 546         move_entries(n, count, n, 0, le32_to_cpu(n->header.nr_entries));
 547 }
 548
 549 /*
 550  * Redistributes entries between two btree nodes to make them
 551  * have similar numbers of entries.
 552  */
 553 static void redistribute2(struct btree_node *left, struct btree_node *right)
 554 {
 555         unsigned int nr_left = le32_to_cpu(left->header.nr_entries);
 556         unsigned int nr_right = le32_to_cpu(right->header.nr_entries);
 557         unsigned int total = nr_left + nr_right;
 558         unsigned int target_left = total / 2;
 559         unsigned int target_right = total - target_left;
 560
 561         if (nr_left < target_left) {
 562                 unsigned int delta = target_left - nr_left;
 563
 564                 copy_entries(left, nr_left, right, 0, delta);
 565                 shift_down(right, delta);
 566         } else if (nr_left > target_left) {
 567                 unsigned int delta = nr_left - target_left;
 568
 569                 if (nr_right)
 570                         shift_up(right, delta);
 571                 copy_entries(right, 0, left, target_left, delta);
 572         }
 573
 574         left->header.nr_entries = cpu_to_le32(target_left);
 575         right->header.nr_entries = cpu_to_le32(target_right);
 576 }
 577
 578 /*
 579  * Redistribute entries between three nodes.  Assumes the central
 580  * node is empty.
 581  */
 582 static void redistribute3(struct btree_node *left, struct btree_node *center,
 583                           struct btree_node *right)
 584 {
 585         unsigned int nr_left = le32_to_cpu(left->header.nr_entries);
 586         unsigned int nr_center = le32_to_cpu(center->header.nr_entries);
 587         unsigned int nr_right = le32_to_cpu(right->header.nr_entries);
 588         unsigned int total, target_left, target_center, target_right;
 589
 590         BUG_ON(nr_center);
 591
 592         total = nr_left + nr_right;
 593         target_left = total / 3;
 594         target_center = (total - target_left) / 2;
 595         target_right = (total - target_left - target_center);
 596
 597         if (nr_left < target_left) {
 598                 unsigned int left_short = target_left - nr_left;
 599
 600                 copy_entries(left, nr_left, right, 0, left_short);
 601                 copy_entries(center, 0, right, left_short, target_center);
 602                 shift_down(right, nr_right - target_right);
 603
 604         } else if (nr_left < (target_left + target_center)) {
 605                 unsigned int left_to_center = nr_left - target_left;
 606
 607                 copy_entries(center, 0, left, target_left, left_to_center);
 608                 copy_entries(center, left_to_center, right, 0, target_center - left_to_center);
 609                 shift_down(right, nr_right - target_right);
 610
 611         } else {
 612                 unsigned int right_short = target_right - nr_right;
 613
 614                 shift_up(right, right_short);
 615                 copy_entries(right, 0, left, nr_left - right_short, right_short);
 616                 copy_entries(center, 0, left, target_left, nr_left - target_left);
 617         }
 618
 619         left->header.nr_entries = cpu_to_le32(target_left);
 620         center->header.nr_entries = cpu_to_le32(target_center);
 621         right->header.nr_entries = cpu_to_le32(target_right);
 622 }
 623
 624 /*
 625  * Splits a node by creating a sibling node and shifting half the nodes
 626  * contents across.  Assumes there is a parent node, and it has room for
 627  * another child.
 628  *
 629  * Before:
 630  *        +--------+
 631  *        | Parent |
 632  *        +--------+
 633  *           |
 634  *           v
 635  *      +----------+
 636  *      | A ++++++ |
 637  *      +----------+
 638  *
 639  *
 640  * After:
 641  *              +--------+
 642  *              | Parent |
 643  *              +--------+
 644  *                |     |
 645  *                v     +------+
 646  *          +---------+        |
 647  *          | A* +++  |        v
 648  *          +---------+   +-------+
 649  *                        | B +++ |
 650  *                        +-------+
 651  *
 652  * Where A* is a shadow of A.
 653  */
 654 static int split_one_into_two(struct shadow_spine *s, unsigned int parent_index,
 655                               struct dm_btree_value_type *vt, uint64_t key)
 656 {
 657         int r;
 658         struct dm_block *left, *right, *parent;
 659         struct btree_node *ln, *rn, *pn;
 660         __le64 location;
 661
 662         left = shadow_current(s);
 663
 664         r = new_block(s->info, &right);
 665         if (r < 0)
 666                 return r;
 667
 668         ln = dm_block_data(left);
 669         rn = dm_block_data(right);
 670
 671         rn->header.flags = ln->header.flags;
 672         rn->header.nr_entries = cpu_to_le32(0);
 673         rn->header.max_entries = ln->header.max_entries;
 674         rn->header.value_size = ln->header.value_size;
 675         redistribute2(ln, rn);
 676
 677         /* patch up the parent */
 678         parent = shadow_parent(s);
 679         pn = dm_block_data(parent);
 680
 681         location = cpu_to_le64(dm_block_location(right));
 682         __dm_bless_for_disk(&location);
 683         r = insert_at(sizeof(__le64), pn, parent_index + 1,
 684                       le64_to_cpu(rn->keys[0]), &location);
 685         if (r) {
 686                 unlock_block(s->info, right);
 687                 return r;
 688         }
 689
 690         /* patch up the spine */
 691         if (key < le64_to_cpu(rn->keys[0])) {
 692                 unlock_block(s->info, right);
 693                 s->nodes[1] = left;
 694         } else {
 695                 unlock_block(s->info, left);
 696                 s->nodes[1] = right;
 697         }
 698
 699         return 0;
 700 }
 701
 702 /*
 703  * We often need to modify a sibling node.  This function shadows a particular
 704  * child of the given parent node.  Making sure to update the parent to point
 705  * to the new shadow.
 706  */
 707 static int shadow_child(struct dm_btree_info *info, struct dm_btree_value_type *vt,
 708                         struct btree_node *parent, unsigned int index,
 709                         struct dm_block **result)
 710 {
 711         int r, inc;
 712         dm_block_t root;
 713         struct btree_node *node;
 714
 715         root = value64(parent, index);
 716
 717         r = dm_tm_shadow_block(info->tm, root, &btree_node_validator,
 718                                result, &inc);
 719         if (r)
 720                 return r;
 721
 722         node = dm_block_data(*result);
 723
 724         if (inc)
 725                 inc_children(info->tm, node, vt);
 726
 727         *((__le64 *) value_ptr(parent, index)) =
 728                 cpu_to_le64(dm_block_location(*result));
 729
 730         return 0;
 731 }
 732
 733 /*
 734  * Splits two nodes into three.  This is more work, but results in fuller
 735  * nodes, so saves metadata space.
 736  */
 737 static int split_two_into_three(struct shadow_spine *s, unsigned int parent_index,
 738                                 struct dm_btree_value_type *vt, uint64_t key)
 739 {
 740         int r;
 741         unsigned int middle_index;
 742         struct dm_block *left, *middle, *right, *parent;
 743         struct btree_node *ln, *rn, *mn, *pn;
 744         __le64 location;
 745
 746         parent = shadow_parent(s);
 747         pn = dm_block_data(parent);
 748
 749         if (parent_index == 0) {
 750                 middle_index = 1;
 751                 left = shadow_current(s);
 752                 r = shadow_child(s->info, vt, pn, parent_index + 1, &right);
 753                 if (r)
 754                         return r;
 755         } else {
 756                 middle_index = parent_index;
 757                 right = shadow_current(s);
 758                 r = shadow_child(s->info, vt, pn, parent_index - 1, &left);
 759                 if (r)
 760                         return r;
 761         }
 762
 763         r = new_block(s->info, &middle);
 764         if (r < 0)
 765                 return r;
 766
 767         ln = dm_block_data(left);
 768         mn = dm_block_data(middle);
 769         rn = dm_block_data(right);
 770
 771         mn->header.nr_entries = cpu_to_le32(0);
 772         mn->header.flags = ln->header.flags;
 773         mn->header.max_entries = ln->header.max_entries;
 774         mn->header.value_size = ln->header.value_size;
 775
 776         redistribute3(ln, mn, rn);
 777
 778         /* patch up the parent */
 779         pn->keys[middle_index] = rn->keys[0];
 780         location = cpu_to_le64(dm_block_location(middle));
 781         __dm_bless_for_disk(&location);
 782         r = insert_at(sizeof(__le64), pn, middle_index,
 783                       le64_to_cpu(mn->keys[0]), &location);
 784         if (r) {
 785                 if (shadow_current(s) != left)
 786                         unlock_block(s->info, left);
 787
 788                 unlock_block(s->info, middle);
 789
 790                 if (shadow_current(s) != right)
 791                         unlock_block(s->info, right);
 792
 793                 return r;
 794         }
 795
 796
 797         /* patch up the spine */
 798         if (key < le64_to_cpu(mn->keys[0])) {
 799                 unlock_block(s->info, middle);
 800                 unlock_block(s->info, right);
 801                 s->nodes[1] = left;
 802         } else if (key < le64_to_cpu(rn->keys[0])) {
 803                 unlock_block(s->info, left);
 804                 unlock_block(s->info, right);
 805                 s->nodes[1] = middle;
 806         } else {
 807                 unlock_block(s->info, left);
 808                 unlock_block(s->info, middle);
 809                 s->nodes[1] = right;
 810         }
 811
 812         return 0;
 813 }
 814
 815 /*----------------------------------------------------------------*/
 816
 817 /*
 818  * Splits a node by creating two new children beneath the given node.
 819  *
 820  * Before:
 821  *        +----------+
 822  *        | A ++++++ |
 823  *        +----------+
 824  *
 825  *
 826  * After:
 827  *      +------------+
 828  *      | A (shadow) |
 829  *      +------------+
 830  *          |   |
 831  *   +------+   +----+
 832  *   |               |
 833  *   v               v
 834  * +-------+     +-------+
 835  * | B +++ |     | C +++ |
 836  * +-------+     +-------+
 837  */
 838 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
 839 {
 840         int r;
 841         size_t size;
 842         unsigned int nr_left, nr_right;
 843         struct dm_block *left, *right, *new_parent;
 844         struct btree_node *pn, *ln, *rn;
 845         __le64 val;
 846
 847         new_parent = shadow_current(s);
 848
 849         pn = dm_block_data(new_parent);
 850         size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
 851                 sizeof(__le64) : s->info->value_type.size;
 852
 853         /* create & init the left block */
 854         r = new_block(s->info, &left);
 855         if (r < 0)
 856                 return r;
 857
 858         ln = dm_block_data(left);
 859         nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
 860
 861         ln->header.flags = pn->header.flags;
 862         ln->header.nr_entries = cpu_to_le32(nr_left);
 863         ln->header.max_entries = pn->header.max_entries;
 864         ln->header.value_size = pn->header.value_size;
 865         memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
 866         memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
 867
 868         /* create & init the right block */
 869         r = new_block(s->info, &right);
 870         if (r < 0) {
 871                 unlock_block(s->info, left);
 872                 return r;
 873         }
 874
 875         rn = dm_block_data(right);
 876         nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
 877
 878         rn->header.flags = pn->header.flags;
 879         rn->header.nr_entries = cpu_to_le32(nr_right);
 880         rn->header.max_entries = pn->header.max_entries;
 881         rn->header.value_size = pn->header.value_size;
 882         memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
 883         memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
 884                nr_right * size);
 885
 886         /* new_parent should just point to l and r now */
 887         pn->header.flags = cpu_to_le32(INTERNAL_NODE);
 888         pn->header.nr_entries = cpu_to_le32(2);
 889         pn->header.max_entries = cpu_to_le32(
 890                 calc_max_entries(sizeof(__le64),
 891                                  dm_bm_block_size(
 892                                          dm_tm_get_bm(s->info->tm))));
 893         pn->header.value_size = cpu_to_le32(sizeof(__le64));
 894
 895         val = cpu_to_le64(dm_block_location(left));
 896         __dm_bless_for_disk(&val);
 897         pn->keys[0] = ln->keys[0];
 898         memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
 899
 900         val = cpu_to_le64(dm_block_location(right));
 901         __dm_bless_for_disk(&val);
 902         pn->keys[1] = rn->keys[0];
 903         memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
 904
 905         unlock_block(s->info, left);
 906         unlock_block(s->info, right);
 907         return 0;
 908 }
 909
 910 /*----------------------------------------------------------------*/
 911
 912 /*
 913  * Redistributes a node's entries with its left sibling.
 914  */
 915 static int rebalance_left(struct shadow_spine *s, struct dm_btree_value_type *vt,
 916                           unsigned int parent_index, uint64_t key)
 917 {
 918         int r;
 919         struct dm_block *sib;
 920         struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
 921
 922         r = shadow_child(s->info, vt, parent, parent_index - 1, &sib);
 923         if (r)
 924                 return r;
 925
 926         left = dm_block_data(sib);
 927         right = dm_block_data(shadow_current(s));
 928         redistribute2(left, right);
 929         *key_ptr(parent, parent_index) = right->keys[0];
 930
 931         if (key < le64_to_cpu(right->keys[0])) {
 932                 unlock_block(s->info, s->nodes[1]);
 933                 s->nodes[1] = sib;
 934         } else {
 935                 unlock_block(s->info, sib);
 936         }
 937
 938         return 0;
 939 }
 940
 941 /*
 942  * Redistributes a nodes entries with its right sibling.
 943  */
 944 static int rebalance_right(struct shadow_spine *s, struct dm_btree_value_type *vt,
 945                            unsigned int parent_index, uint64_t key)
 946 {
 947         int r;
 948         struct dm_block *sib;
 949         struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
 950
 951         r = shadow_child(s->info, vt, parent, parent_index + 1, &sib);
 952         if (r)
 953                 return r;
 954
 955         left = dm_block_data(shadow_current(s));
 956         right = dm_block_data(sib);
 957         redistribute2(left, right);
 958         *key_ptr(parent, parent_index + 1) = right->keys[0];
 959
 960         if (key < le64_to_cpu(right->keys[0])) {
 961                 unlock_block(s->info, sib);
 962         } else {
 963                 unlock_block(s->info, s->nodes[1]);
 964                 s->nodes[1] = sib;
 965         }
 966
 967         return 0;
 968 }
 969
 970 /*
 971  * Returns the number of spare entries in a node.
 972  */
 973 static int get_node_free_space(struct dm_btree_info *info, dm_block_t b, unsigned int *space)
 974 {
 975         int r;
 976         unsigned int nr_entries;
 977         struct dm_block *block;
 978         struct btree_node *node;
 979
 980         r = bn_read_lock(info, b, &block);
 981         if (r)
 982                 return r;
 983
 984         node = dm_block_data(block);
 985         nr_entries = le32_to_cpu(node->header.nr_entries);
 986         *space = le32_to_cpu(node->header.max_entries) - nr_entries;
 987
 988         unlock_block(info, block);
 989         return 0;
 990 }
 991
 992 /*
 993  * Make space in a node, either by moving some entries to a sibling,
 994  * or creating a new sibling node.  SPACE_THRESHOLD defines the minimum
 995  * number of free entries that must be in the sibling to make the move
 996  * worth while.  If the siblings are shared (eg, part of a snapshot),
 997  * then they are not touched, since this break sharing and so consume
 998  * more space than we save.
 999  */
1000 #define SPACE_THRESHOLD 8
1001 static int rebalance_or_split(struct shadow_spine *s, struct dm_btree_value_type *vt,
1002                               unsigned int parent_index, uint64_t key)
1003 {
1004         int r;
1005         struct btree_node *parent = dm_block_data(shadow_parent(s));
1006         unsigned int nr_parent = le32_to_cpu(parent->header.nr_entries);
1007         unsigned int free_space;
1008         int left_shared = 0, right_shared = 0;
1009
1010         /* Should we move entries to the left sibling? */
1011         if (parent_index > 0) {
1012                 dm_block_t left_b = value64(parent, parent_index - 1);
1013
1014                 r = dm_tm_block_is_shared(s->info->tm, left_b, &left_shared);
1015                 if (r)
1016                         return r;
1017
1018                 if (!left_shared) {
1019                         r = get_node_free_space(s->info, left_b, &free_space);
1020                         if (r)
1021                                 return r;
1022
1023                         if (free_space >= SPACE_THRESHOLD)
1024                                 return rebalance_left(s, vt, parent_index, key);
1025                 }
1026         }
1027
1028         /* Should we move entries to the right sibling? */
1029         if (parent_index < (nr_parent - 1)) {
1030                 dm_block_t right_b = value64(parent, parent_index + 1);
1031
1032                 r = dm_tm_block_is_shared(s->info->tm, right_b, &right_shared);
1033                 if (r)
1034                         return r;
1035
1036                 if (!right_shared) {
1037                         r = get_node_free_space(s->info, right_b, &free_space);
1038                         if (r)
1039                                 return r;
1040
1041                         if (free_space >= SPACE_THRESHOLD)
1042                                 return rebalance_right(s, vt, parent_index, key);
1043                 }
1044         }
1045
1046         /*
1047          * We need to split the node, normally we split two nodes
1048          * into three.  But when inserting a sequence that is either
1049          * monotonically increasing or decreasing it's better to split
1050          * a single node into two.
1051          */
1052         if (left_shared || right_shared || (nr_parent <= 2) ||
1053             (parent_index == 0) || (parent_index + 1 == nr_parent)) {
1054                 return split_one_into_two(s, parent_index, vt, key);
1055         } else {
1056                 return split_two_into_three(s, parent_index, vt, key);
1057         }
1058 }
1059
1060 /*
1061  * Does the node contain a particular key?
1062  */
1063 static bool contains_key(struct btree_node *node, uint64_t key)
1064 {
1065         int i = lower_bound(node, key);
1066
1067         if (i >= 0 && le64_to_cpu(node->keys[i]) == key)
1068                 return true;
1069
1070         return false;
1071 }
1072
1073 /*
1074  * In general we preemptively make sure there's a free entry in every
1075  * node on the spine when doing an insert.  But we can avoid that with
1076  * leaf nodes if we know it's an overwrite.
1077  */
1078 static bool has_space_for_insert(struct btree_node *node, uint64_t key)
1079 {
1080         if (node->header.nr_entries == node->header.max_entries) {
1081                 if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1082                         /* we don't need space if it's an overwrite */
1083                         return contains_key(node, key);
1084                 }
1085
1086                 return false;
1087         }
1088
1089         return true;
1090 }
1091
1092 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
1093                             struct dm_btree_value_type *vt,
1094                             uint64_t key, unsigned int *index)
1095 {
1096         int r, i = *index, top = 1;
1097         struct btree_node *node;
1098
1099         for (;;) {
1100                 r = shadow_step(s, root, vt);
1101                 if (r < 0)
1102                         return r;
1103
1104                 node = dm_block_data(shadow_current(s));
1105
1106                 /*
1107                  * We have to patch up the parent node, ugly, but I don't
1108                  * see a way to do this automatically as part of the spine
1109                  * op.
1110                  */
1111                 if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
1112                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1113
1114                         __dm_bless_for_disk(&location);
1115                         memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1116                                     &location, sizeof(__le64));
1117                 }
1118
1119                 node = dm_block_data(shadow_current(s));
1120
1121                 if (!has_space_for_insert(node, key)) {
1122                         if (top)
1123                                 r = btree_split_beneath(s, key);
1124                         else
1125                                 r = rebalance_or_split(s, vt, i, key);
1126
1127                         if (r < 0)
1128                                 return r;
1129
1130                         /* making space can cause the current node to change */
1131                         node = dm_block_data(shadow_current(s));
1132                 }
1133
1134                 i = lower_bound(node, key);
1135
1136                 if (le32_to_cpu(node->header.flags) & LEAF_NODE)
1137                         break;
1138
1139                 if (i < 0) {
1140                         /* change the bounds on the lowest key */
1141                         node->keys[0] = cpu_to_le64(key);
1142                         i = 0;
1143                 }
1144
1145                 root = value64(node, i);
1146                 top = 0;
1147         }
1148
1149         if (i < 0 || le64_to_cpu(node->keys[i]) != key)
1150                 i++;
1151
1152         *index = i;
1153         return 0;
1154 }
1155
1156 static int __btree_get_overwrite_leaf(struct shadow_spine *s, dm_block_t root,
1157                                       uint64_t key, int *index)
1158 {
1159         int r, i = -1;
1160         struct btree_node *node;
1161
1162         *index = 0;
1163         for (;;) {
1164                 r = shadow_step(s, root, &s->info->value_type);
1165                 if (r < 0)
1166                         return r;
1167
1168                 node = dm_block_data(shadow_current(s));
1169
1170                 /*
1171                  * We have to patch up the parent node, ugly, but I don't
1172                  * see a way to do this automatically as part of the spine
1173                  * op.
1174                  */
1175                 if (shadow_has_parent(s) && i >= 0) {
1176                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1177
1178                         __dm_bless_for_disk(&location);
1179                         memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1180                                     &location, sizeof(__le64));
1181                 }
1182
1183                 node = dm_block_data(shadow_current(s));
1184                 i = lower_bound(node, key);
1185
1186                 BUG_ON(i < 0);
1187                 BUG_ON(i >= le32_to_cpu(node->header.nr_entries));
1188
1189                 if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1190                         if (key != le64_to_cpu(node->keys[i]))
1191                                 return -EINVAL;
1192                         break;
1193                 }
1194
1195                 root = value64(node, i);
1196         }
1197
1198         *index = i;
1199         return 0;
1200 }
1201
1202 int btree_get_overwrite_leaf(struct dm_btree_info *info, dm_block_t root,
1203                              uint64_t key, int *index,
1204                              dm_block_t *new_root, struct dm_block **leaf)
1205 {
1206         int r;
1207         struct shadow_spine spine;
1208
1209         BUG_ON(info->levels > 1);
1210         init_shadow_spine(&spine, info);
1211         r = __btree_get_overwrite_leaf(&spine, root, key, index);
1212         if (!r) {
1213                 *new_root = shadow_root(&spine);
1214                 *leaf = shadow_current(&spine);
1215
1216                 /*
1217                  * Decrement the count so exit_shadow_spine() doesn't
1218                  * unlock the leaf.
1219                  */
1220                 spine.count--;
1221         }
1222         exit_shadow_spine(&spine);
1223
1224         return r;
1225 }
1226
1227 static bool need_insert(struct btree_node *node, uint64_t *keys,
1228                         unsigned int level, unsigned int index)
1229 {
1230         return ((index >= le32_to_cpu(node->header.nr_entries)) ||
1231                 (le64_to_cpu(node->keys[index]) != keys[level]));
1232 }
1233
1234 static int insert(struct dm_btree_info *info, dm_block_t root,
1235                   uint64_t *keys, void *value, dm_block_t *new_root,
1236                   int *inserted)
1237                   __dm_written_to_disk(value)
1238 {
1239         int r;
1240         unsigned int level, index = -1, last_level = info->levels - 1;
1241         dm_block_t block = root;
1242         struct shadow_spine spine;
1243         struct btree_node *n;
1244         struct dm_btree_value_type le64_type;
1245
1246         init_le64_type(info->tm, &le64_type);
1247         init_shadow_spine(&spine, info);
1248
1249         for (level = 0; level < (info->levels - 1); level++) {
1250                 r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
1251                 if (r < 0)
1252                         goto bad;
1253
1254                 n = dm_block_data(shadow_current(&spine));
1255
1256                 if (need_insert(n, keys, level, index)) {
1257                         dm_block_t new_tree;
1258                         __le64 new_le;
1259
1260                         r = dm_btree_empty(info, &new_tree);
1261                         if (r < 0)
1262                                 goto bad;
1263
1264                         new_le = cpu_to_le64(new_tree);
1265                         __dm_bless_for_disk(&new_le);
1266
1267                         r = insert_at(sizeof(uint64_t), n, index,
1268                                       keys[level], &new_le);
1269                         if (r)
1270                                 goto bad;
1271                 }
1272
1273                 if (level < last_level)
1274                         block = value64(n, index);
1275         }
1276
1277         r = btree_insert_raw(&spine, block, &info->value_type,
1278                              keys[level], &index);
1279         if (r < 0)
1280                 goto bad;
1281
1282         n = dm_block_data(shadow_current(&spine));
1283
1284         if (need_insert(n, keys, level, index)) {
1285                 if (inserted)
1286                         *inserted = 1;
1287
1288                 r = insert_at(info->value_type.size, n, index,
1289                               keys[level], value);
1290                 if (r)
1291                         goto bad_unblessed;
1292         } else {
1293                 if (inserted)
1294                         *inserted = 0;
1295
1296                 if (info->value_type.dec &&
1297                     (!info->value_type.equal ||
1298                      !info->value_type.equal(
1299                              info->value_type.context,
1300                              value_ptr(n, index),
1301                              value))) {
1302                         info->value_type.dec(info->value_type.context,
1303                                              value_ptr(n, index), 1);
1304                 }
1305                 memcpy_disk(value_ptr(n, index),
1306                             value, info->value_type.size);
1307         }
1308
1309         *new_root = shadow_root(&spine);
1310         exit_shadow_spine(&spine);
1311
1312         return 0;
1313
1314 bad:
1315         __dm_unbless_for_disk(value);
1316 bad_unblessed:
1317         exit_shadow_spine(&spine);
1318         return r;
1319 }
1320
1321 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
1322                     uint64_t *keys, void *value, dm_block_t *new_root)
1323         __dm_written_to_disk(value)
1324 {
1325         return insert(info, root, keys, value, new_root, NULL);
1326 }
1327 EXPORT_SYMBOL_GPL(dm_btree_insert);
1328
1329 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
1330                            uint64_t *keys, void *value, dm_block_t *new_root,
1331                            int *inserted)
1332         __dm_written_to_disk(value)
1333 {
1334         return insert(info, root, keys, value, new_root, inserted);
1335 }
1336 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
1337
1338 /*----------------------------------------------------------------*/
1339
1340 static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
1341                     uint64_t *result_key, dm_block_t *next_block)
1342 {
1343         int i, r;
1344         uint32_t flags;
1345
1346         do {
1347                 r = ro_step(s, block);
1348                 if (r < 0)
1349                         return r;
1350
1351                 flags = le32_to_cpu(ro_node(s)->header.flags);
1352                 i = le32_to_cpu(ro_node(s)->header.nr_entries);
1353                 if (!i)
1354                         return -ENODATA;
1355
1356                 i--;
1357
1358                 if (find_highest)
1359                         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
1360                 else
1361                         *result_key = le64_to_cpu(ro_node(s)->keys[0]);
1362
1363                 if (next_block || flags & INTERNAL_NODE) {
1364                         if (find_highest)
1365                                 block = value64(ro_node(s), i);
1366                         else
1367                                 block = value64(ro_node(s), 0);
1368                 }
1369
1370         } while (flags & INTERNAL_NODE);
1371
1372         if (next_block)
1373                 *next_block = block;
1374         return 0;
1375 }
1376
1377 static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
1378                              bool find_highest, uint64_t *result_keys)
1379 {
1380         int r = 0, count = 0, level;
1381         struct ro_spine spine;
1382
1383         init_ro_spine(&spine, info);
1384         for (level = 0; level < info->levels; level++) {
1385                 r = find_key(&spine, root, find_highest, result_keys + level,
1386                              level == info->levels - 1 ? NULL : &root);
1387                 if (r == -ENODATA) {
1388                         r = 0;
1389                         break;
1390
1391                 } else if (r)
1392                         break;
1393
1394                 count++;
1395         }
1396         exit_ro_spine(&spine);
1397
1398         return r ? r : count;
1399 }
1400
1401 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
1402                               uint64_t *result_keys)
1403 {
1404         return dm_btree_find_key(info, root, true, result_keys);
1405 }
1406 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
1407
1408 int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
1409                              uint64_t *result_keys)
1410 {
1411         return dm_btree_find_key(info, root, false, result_keys);
1412 }
1413 EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
1414
1415 /*----------------------------------------------------------------*/
1416
1417 /*
1418  * FIXME: We shouldn't use a recursive algorithm when we have limited stack
1419  * space.  Also this only works for single level trees.
1420  */
1421 static int walk_node(struct dm_btree_info *info, dm_block_t block,
1422                      int (*fn)(void *context, uint64_t *keys, void *leaf),
1423                      void *context)
1424 {
1425         int r;
1426         unsigned int i, nr;
1427         struct dm_block *node;
1428         struct btree_node *n;
1429         uint64_t keys;
1430
1431         r = bn_read_lock(info, block, &node);
1432         if (r)
1433                 return r;
1434
1435         n = dm_block_data(node);
1436
1437         nr = le32_to_cpu(n->header.nr_entries);
1438         for (i = 0; i < nr; i++) {
1439                 if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
1440                         r = walk_node(info, value64(n, i), fn, context);
1441                         if (r)
1442                                 goto out;
1443                 } else {
1444                         keys = le64_to_cpu(*key_ptr(n, i));
1445                         r = fn(context, &keys, value_ptr(n, i));
1446                         if (r)
1447                                 goto out;
1448                 }
1449         }
1450
1451 out:
1452         dm_tm_unlock(info->tm, node);
1453         return r;
1454 }
1455
1456 int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
1457                   int (*fn)(void *context, uint64_t *keys, void *leaf),
1458                   void *context)
1459 {
1460         BUG_ON(info->levels > 1);
1461         return walk_node(info, root, fn, context);
1462 }
1463 EXPORT_SYMBOL_GPL(dm_btree_walk);
1464
1465 /*----------------------------------------------------------------*/
1466
1467 static void prefetch_values(struct dm_btree_cursor *c)
1468 {
1469         unsigned int i, nr;
1470         __le64 value_le;
1471         struct cursor_node *n = c->nodes + c->depth - 1;
1472         struct btree_node *bn = dm_block_data(n->b);
1473         struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
1474
1475         BUG_ON(c->info->value_type.size != sizeof(value_le));
1476
1477         nr = le32_to_cpu(bn->header.nr_entries);
1478         for (i = 0; i < nr; i++) {
1479                 memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1480                 dm_bm_prefetch(bm, le64_to_cpu(value_le));
1481         }
1482 }
1483
1484 static bool leaf_node(struct dm_btree_cursor *c)
1485 {
1486         struct cursor_node *n = c->nodes + c->depth - 1;
1487         struct btree_node *bn = dm_block_data(n->b);
1488
1489         return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1490 }
1491
1492 static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1493 {
1494         int r;
1495         struct cursor_node *n = c->nodes + c->depth;
1496
1497         if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1498                 DMERR("couldn't push cursor node, stack depth too high");
1499                 return -EINVAL;
1500         }
1501
1502         r = bn_read_lock(c->info, b, &n->b);
1503         if (r)
1504                 return r;
1505
1506         n->index = 0;
1507         c->depth++;
1508
1509         if (c->prefetch_leaves || !leaf_node(c))
1510                 prefetch_values(c);
1511
1512         return 0;
1513 }
1514
1515 static void pop_node(struct dm_btree_cursor *c)
1516 {
1517         c->depth--;
1518         unlock_block(c->info, c->nodes[c->depth].b);
1519 }
1520
1521 static int inc_or_backtrack(struct dm_btree_cursor *c)
1522 {
1523         struct cursor_node *n;
1524         struct btree_node *bn;
1525
1526         for (;;) {
1527                 if (!c->depth)
1528                         return -ENODATA;
1529
1530                 n = c->nodes + c->depth - 1;
1531                 bn = dm_block_data(n->b);
1532
1533                 n->index++;
1534                 if (n->index < le32_to_cpu(bn->header.nr_entries))
1535                         break;
1536
1537                 pop_node(c);
1538         }
1539
1540         return 0;
1541 }
1542
1543 static int find_leaf(struct dm_btree_cursor *c)
1544 {
1545         int r = 0;
1546         struct cursor_node *n;
1547         struct btree_node *bn;
1548         __le64 value_le;
1549
1550         for (;;) {
1551                 n = c->nodes + c->depth - 1;
1552                 bn = dm_block_data(n->b);
1553
1554                 if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1555                         break;
1556
1557                 memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1558                 r = push_node(c, le64_to_cpu(value_le));
1559                 if (r) {
1560                         DMERR("push_node failed");
1561                         break;
1562                 }
1563         }
1564
1565         if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1566                 return -ENODATA;
1567
1568         return r;
1569 }
1570
1571 int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1572                           bool prefetch_leaves, struct dm_btree_cursor *c)
1573 {
1574         int r;
1575
1576         c->info = info;
1577         c->root = root;
1578         c->depth = 0;
1579         c->prefetch_leaves = prefetch_leaves;
1580
1581         r = push_node(c, root);
1582         if (r)
1583                 return r;
1584
1585         return find_leaf(c);
1586 }
1587 EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1588
1589 void dm_btree_cursor_end(struct dm_btree_cursor *c)
1590 {
1591         while (c->depth)
1592                 pop_node(c);
1593 }
1594 EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1595
1596 int dm_btree_cursor_next(struct dm_btree_cursor *c)
1597 {
1598         int r = inc_or_backtrack(c);
1599
1600         if (!r) {
1601                 r = find_leaf(c);
1602                 if (r)
1603                         DMERR("find_leaf failed");
1604         }
1605
1606         return r;
1607 }
1608 EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1609
1610 int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count)
1611 {
1612         int r = 0;
1613
1614         while (count-- && !r)
1615                 r = dm_btree_cursor_next(c);
1616
1617         return r;
1618 }
1619 EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1620
1621 int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1622 {
1623         if (c->depth) {
1624                 struct cursor_node *n = c->nodes + c->depth - 1;
1625                 struct btree_node *bn = dm_block_data(n->b);
1626
1627                 if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1628                         return -EINVAL;
1629
1630                 *key = le64_to_cpu(*key_ptr(bn, n->index));
1631                 memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1632                 return 0;
1633
1634         } else
1635                 return -ENODATA;
1636 }
1637 EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);